Machine learning approach for 2D abrasion mapping in Sediment Bypass Tunnels: a case study of Koshibu SBT, Japan

Abstract

Sediment Bypass Tunnels (SBTs) effectively mitigate reservoir sedimentation by diverting flood-laden flows, but they face significant challenges due to hydroabrasive erosion, which compromises their sustainability. Predicting this abrasion is complex due to the intricate interactions between flow hydraulics and sediment transport, along with limited high-quality data. In this study, we explore, for the first time, the potential of using the XGBoost machine learning algorithm to predict the spatial abrasion of SBTs. The Koshibu SBT in Japan, extending approximately 4 km, was selected as the case study. Three experimental scenarios were evaluated: the entire tunnel, the straight section, and the curved section. A spatial abrasion topography was measured using laser scanning tools with a spatial resolution of 2 cm. The controlling factors for abrasion were developed based on geometric and hydraulic features. The abrasion inventory map, consisting of over 1 million data points indicating damaged and non-damaged sites, was divided equally for training and testing the XGBoost algorithm. Results indicate that the XGBoost model effectively predicts 2D spatial abrasions in SBTs, achieving an overall accuracy of 0.864, exceeding 0.9 in some sections. The developed abrasion map accurately captures various complex patterns throughout the tunnel but has some limitations in areas with small wave-like patterns. Overall, this study demonstrates the potential of machine learning algorithms for predicting tunnel abrasion in SBTs.Paper highlightsThis study introduces a validated 2D model for tunnel abrasion based on field data, contributing to improved sediment management in SBTs.ASM Model efficiently predicts abrasion mapping in SBT, achieving 86.4% overall accuracy.High sensitivity and specificity in distinguishing abraded and non-abraded areas.Captures four complex abrasion patterns in straight and curved sections but is limited to relatively small wave-like patterns.Geometric and hydraulic parameters, particularly the elongated distance and flow velocity, exhibit significant impacts in the ASM model.

Keywords

• Spatial abrasion of SBTs
• abrasion inventory map
• XGBoost machine learning
• abrasion pattern predicting